Employing Machine Learning to investigate properties of molecules and periodic systems
||Employing Machine Learning to investigate properties of molecules and periodic systems|
||SNIC Small Compute|
||Rodrigo Pereira de Carvalho <email@example.com>|
||2021-11-05 – 2022-12-01|
Nowadays, it is a common sense on science that the interplay of Artificial Intelligence (more specifically by means of machine learning techniques) and first-principles calculations can lead us to a breakthrough on materials design. Recent works have been employing machine learning to show the possibility to discover new materials , calculate physical properties without the high computational cost of prior techniques  and even to by-pass the solution of Quantum Mechanic equations . For molecular systems, we can produce high-accurate fingerprints to serve as inputs for machine learning, but for periodic systems (solids, surfaces, etc) we face some difficulties, besides new approaches has been developed [2,4]. The specific goals for this project are:
(I) Apply the machine learning machinery to molecular systems to speed-up or by-pass ab initio calculations.
(ii) Benchmark a few options to represent periodic systems (especially aimed as electrodes).
(iii) Interplay of our machinery with DFT to investigate properties of selected materials.
To this date, we have been developing an original database of organic energy materials based on high-quality DFT calculations for molecular and periodic systems. So far, we have an impressive number of ~36000 unique molecular structures, that required more than 310000 DFT calculations, and approximately 30 crystals representing electroactive compounds for Li-ion batteries. From these crystals, we published a paper  in a prestigious journal (ChemSusChem) showing how to tailor organic materials to achieve higher voltages and energy densities in organic batteries. Furthermore, there is a manuscript in its final steps showing an innovative methodology based on Artificial Intelligence (AI) to discover novel organic materials, in which we used the database to analyse 20 million new molecules in a high-throughput approach to propose new cathodes for batteries. These results have recently been published in the prestigious Energy Storage Materials journal . The developed database and the AI-software are being distributed under an open-source basis at GitLab (gitlab.com) and connected with the published paper. Finally, and connected with all the developed data and published results, we are expanding our investigation of redox materials for batteries to an important aspect: stability. The goal is to use another AI-based method to identify the redox stability of a given compound, predicting if it would be of interesting for battery applications.
1. Balachandran, Prasanna V., et al. Physical Review Materials 2.4 (2018)
2. Faber, Felix, et al. International Journal of Quantum Chemistry 115.16 (2015)
3. Brockherde, Felix, et al. Nature communications 8.1 (2017)
4. Huo, Haoyan, and Matthias Rupp. "Unified representation for machine learning of molecules and crystals." preprint (2017).
5. R. P. Carvalho, C. F. N. Marchiori, D. Brandell, C. M. Araujo, ChemSusChem 2020, 13, 2402.
6. R. P. Carvalho, C. F. N. Marchiori, D. Brandell, C. M. Araujo, Energy Storage Materials 2021.